[WIP] TDB traverse lock changes for massive AD DC perf improvement

[Garming Sam]

I didn't expect that the performance improvement would be so large. It
goes to show that there was never any reason for unindexed searches to
be orders of magnitude slower. The fact that unindexed and indexed
searches are now reasonably comparable shows that we probably do need to
be working on our indexing schemes.


On 01/04/17 18:44, Andrew Bartlett via samba-technical wrote:
> On Fri, 2017-03-31 at 13:23 -0700, Jeremy Allison wrote:
>> On Sat, Apr 01, 2017 at 09:05:48AM +1300, Andrew Bartlett via samba-
>> technical wrote:
>>> I've now done that.  This feels closer to the correct solution, but
>>> throws solaris10 under the bus for now.  See the attached patch.
>> Solaris10 is already under the bus wheels. They incorrectly
>> prevent an existing open UNIX domain socket from being accessed
>> from a non-root process (i.e. their permission check is being
>> done on time of use, not time of open). This breaks asynchronous
>> messaging in Samba.
>>
>> I've logged this as a bug with Oracle, but I don't expect
>> this will ever get fixed.
>>
>> It does mean that we should probably stop caring about
>> Solaris long term and put it in the same bucket as HPUX/IRIX/
>> AIX/True64 and the like. Sad, but practical.
> Thanks.  I wasn't at all sure how to translate the few details in the
> commit message back into a test case and then workaround.
>
> This set of patches passes a full make test for me 3 times over (it
> wouldn't shock me if this fixes some flapping tests). 
>
> Graphs of the perf tests with these patches attached.  The performance
> improvement for un-indexed searches on the AD DC is really impressive!
>
> Aside from a good delay so others can comment, what would you recommend
> I do next towards progressing this patch?
>
> Thanks,
>
> Andrew Bartlett
>